Subscribe
TEST RECORD · ← all field reports

FIELD REPORTS/IMAGE/ISSUE #123

Paints images pixel by pixel from a prompt, and the same model also estimates depth and restores blurry photos

Almost every image generator you have used works on a compressed sketch of the picture and hands the final pixels to a separate decoder. Iris-3B skips that step and paints the pixels itself. The surprise is what else the same model learned to do along the way.

MODELIris-3B
PUBLISHEDOctober 12, 2026
READ TIME3 min
CATEGORYIMAGE

Field notes

01 What it does

You type a plain English sentence and get back an image of about one megapixel, at the aspect ratio you pick. So far it looks like any other image generator. The difference is in how the picture is made: most generators work on a compressed version of the image and rely on a separate decoder to turn it into pixels, and that is where fine texture tends to get smoothed away. Iris-3B outputs every pixel itself, so skin, fabric and stone keep their grain.

The second half is the interesting part. The team took the same 3 billion parameter model, fine-tuned it without changing its design, and it turned into a general vision model. Give it a photo and it returns a depth map, showing what is near and what is far. Give it a small or degraded photo and it returns a restored version at four times the size. Both run from a single picture in one pass, no prompt needed.

02 How to try it

The Iris-3B demo Space runs in your browser on a shared GPU and has three tabs: Generate, Depth and Upscale. Start in Generate with a portrait prompt that asks for skin texture or film grain, then zoom into the result: the grain should be real detail, not blur. Next open Upscale, drop in a small photo from your phone and compare the result with the original side by side. Running it yourself needs an NVIDIA graphics card and a 12 GB download for image generation; the depth and upscaling weights are a separate 12 GB each.

03 Caveat

Generation is slow by current standards, around 100 steps for a full quality image, and text inside images and exact object counts are unreliable. Depth is relative rather than measured in meters, so it gives you ordering and shape, not distances.

04 What you can do with it

  • Generate portraits and product shots where fine texture matters.
  • Turn a single photo into a depth map for parallax effects, relighting or 3D mockups.
  • Enlarge old or low resolution photos before printing them.
  • Clean up compressed images pulled from the web or an old archive.
  • Feed it a text file of prompts and generate the whole batch in one run.

Try the demo

View model page

Read this issue on neuralexpedition.com →